← Back ICRA 2026

Continual Learning for Traversability Prediction with Uncertainty-Aware Adaptation

PDF

AI summary

Key figure (auto-extracted from paper)

Robots can continuously adapt to new terrains without catastrophic forgetting or storing historical data.

Continual learning Traversability prediction Uncertainty-aware adaptation Generative experience recall Autonomous navigation Field robots

Problem

Learning-based traversability predictors struggle to adapt to novel terrains and suffer from catastrophic forgetting, while traditional experience replay methods demand excessive memory or risk data loss.

Approach

The framework uses a generative model to synthesize past training samples on-the-fly and filters them using uncertainty estimates, enabling reliable model updates without retaining raw historical data.

Key results

Uncertainty-aware traversability prediction linking terrain features to robot dynamics
Generative experience recall framework preventing catastrophic forgetting
Real-world validation on a skid-steering robot across diverse terrains
Memory-efficient adaptation via on-the-fly data synthesis

Why it matters

Enables scalable, memory-efficient continual adaptation for autonomous robots navigating unstructured and changing environments.

Abstract

Traversability prediction is a critical component of autonomous navigation in unstructured environments, where com- plex and uncertain robot-terrain interactions pose significant chal- lenges such as traction loss and dynamic instability. Despite recent progress in learning-based traversability prediction, these meth- ods often fail to adapt to novel terrains. Even when adaptation is achieved, retaining experience from previously trained envi- ronments remains a challenge, a problem known as catastrophic forgetting. To address this challenge, we propose a continual learn- ing framework for traversability prediction that incrementally adapts to new terrains using a generative experience recall model. A key virtue of the proposed framework is two folds: i) retain prior experience without storing past data; and ii) incorporate the uncertainty of the generated samples from the recall model, enabling uncertainty-aware adaptation. Real-world experiments with a skid-steering robot validate the effectiveness of the proposed framework, demonstrating its ability to adapt across a series of diverse environments while mitigating catastrophic forgetting.

Index terms

Continual Learning Planning under Uncertainty Field Robots